AMD AI director Stella Laurenzo dropped a performance audit directly on Claude Code's official GitHub repository, backed by logs from 6,852 real engineering sessions. The numbers are brutal: median thinking depth plunged 67% after a major February update, pre-coding research effort fell about 70%, and the model displayed "dodging" or "early termination" behavior 173 times in 17 days.
The cost spike hit hardest. For the same workload, the monthly API bill rocketed from $345 to $42,121 — a 122x jump. Laurenzo concluded that "Claude can no longer be trusted for complex engineering tasks," and her team has already switched to other providers. She also flagged that huge chunks of tokens (API billing units) were burned on unproductive loops rather than real work.
Anthropic: It's a Parameter Tweak, Not Degradation
Anthropic's Boris replied on GitHub, attributing the issues to two parameter changes: the February 9 introduction of an "adaptive thinking" mechanism, and the March 3 drop of the default thinking level from "high" to "medium." He stressed the model's core capability hasn't degraded, and users can restore performance by manually setting the effort parameter to maximum.
Developers aren't convinced. Many report that even at max effort, the model still rushes through tasks and delivers shallow answers — far more than before. They feel the official explanation misses the point.
Hidden Thinking Content Fuels Distrust
Laurenzo's report highlighted another sore point: Anthropic's earlier launch of a "hidden thinking content" feature, which obscures the model's full reasoning chain. In normal operation it's just a UI choice, but during regression it hides the very clues that would let users spot the model slacking off.
Since the report surfaced, many Claude Code subscribers have canceled and moved to OpenAI Codex or other tools. Several developers say they felt the performance dip weeks ago but lacked hard data — now they have it.